GLM-5.3 Flash

Active Parameters

313.3B

Context Length

1.05M

Modality

Text

Architecture

Mixture of Experts

KV Cache Rate

1440 KB/tok

License

MIT License

Release Date

25 Aug 2026

Knowledge Cutoff

-

API Pricing (per 1M)

Input: $0.15 · Output: $0.50

Hardware & System Requirements

Precision

Native

Context

Workload Memory: 629.54 GB

FP16 · 8K context

Base Weights: 615.79 GB

KV Cache: 11.25 GB

Runtime & Graphs: 2.50 GB

Minimum Required Hardware

8x NVIDIA H200 141GB

629.5 / 1015.2 GB usable · TP 8

Determine requirements for your specific workload:

Upgrade to Pro

Runtime requirements

Checkpoint format

FP16

CUDA

>= 11.8

PyTorch

>= 2.1

Transformers

>= 5.16.0

ArchitecturePrecisionAttention kernel

Blackwell, Hopper

Native FP16

FlashAttention-3

Ada Lovelace, Ampere

Native FP16

FlashAttention-2

Architecture Diagram

Input TokensToken EmbeddingPosition: RoPEHidden: 4.1k · Context: 1.05M · Vocab: 154.9kx 45 layersRMSNormPre-AttentionDeepSeek Sparse Attention64Q / 64KV headsHead dim: 128+RMSNormPre-FFNSparse MoE FFN (8/288 experts)SwiGLUIntermediate: 2k+Final RMSNormOutput Logits

Evaluation Benchmarks

Rank

#84

BenchmarkScoreRank

Agentic Index

Artificial Analysis

0.51

13

Agentic Coding

LiveBench Agentic

0.57

22

0.79

27

0.76

31

Web Development

WebDev Arena

1610.48

35

0.71

37

General Text

Text Arena

1475

40

Agent Arena

Agent Arena

1045.63

46

0.72

51

LiveBench Average

LiveBench Average

0.72

51

0.81

54

0.78

57

Intelligence Index

Artificial Analysis

0.28

118

Rankings

Overall Rank

#84

Coding Rank

#58

About GLM-5.3 Flash

GLM-5.3 Flash is a high-efficiency open model designed for rapid inference and low-latency reasoning workloads. Built on the Glm5Next architecture, it offers optimized throughput for agentic applications and code generation.

Specifications

Attention

Attention Structure

DeepSeek Sparse Attention

Attention Heads

64

Key-Value Heads

64

Attention Head Dimension

128

Position Embedding

ROPE

Linear Attention

Yes

Linear Attention Ratio

75.6%

Normalization

RMS Normalization

Activation Function

SwigLU

Dimensions

Base Precision

FP16

Hidden Dimension Size

4,096

Number of Layers

45

FFN Intermediate Size (Dense)

12,288

Multi-Token Prediction Heads

1

Tokenizer

Vocabulary Size

154,880

Mixture of Experts

Total Expert Parameters

17.3B

Number of Experts

288

Active Experts

8

Shared Experts

1

FFN Intermediate Size (per Expert)

2,048

Dense Layers Before MoE

3

About GLM Family

General Language Models from Z.ai


Other GLM Family Models
GLM-5.3 Flash - VRAM Requirements, Specs & Benchmarks